Merge pull request fast-pack#103 from pps83/clear-padding

Clear padding in fastpfor, simdfastpfor, simdgroupsimple
rohan-flutterint · Feb 20, 2023 · 024eed0 · 024eed0
2 parents 2ec76c3 + d67c7ab
commit 024eed0
Show file tree

Hide file tree

Showing 3 changed files with 18 additions and 0 deletions.
diff --git a/headers/fastpfor.h b/headers/fastpfor.h
@@ -200,7 +200,10 @@ class FastPForImpl {
         static_cast<uint32_t>(bc - &bytescontainer[0]);
     *(out++) = bytescontainersize;
     memcpy(out, &bytescontainer[0], bytescontainersize);
+    uint8_t* pad8 = (uint8_t*)out + bytescontainersize;
     out += (bytescontainersize + sizeof(uint32_t) - 1) / sizeof(uint32_t);
+    while (pad8 < (uint8_t*)out)
+        *pad8++ = 0; // clear padding bytes
 
     IntType bitmap = 0;
     for (uint32_t k = 2; k <= sizeof(IntType) * 8; ++k) {
@@ -472,7 +475,11 @@ class SimplePFor : public IntegerCODEC {
         static_cast<uint32_t>(bc - &bytescontainer[0]);
     *(out++) = bytescontainersize;
     memcpy(out, &bytescontainer[0], bytescontainersize);
+    uint8_t* pad8 = (uint8_t*)out + bytescontainersize;
     out += (bytescontainersize + sizeof(uint32_t) - 1) / sizeof(uint32_t);
+    while (pad8 < (uint8_t*)out)
+        *pad8++ = 0; // clear padding bytes
+
     size_t outcap = 0;
     ecoder.encodeArray(datatobepacked.data(), datatobepacked.size(), out,
                        outcap);

diff --git a/headers/simdfastpfor.h b/headers/simdfastpfor.h
@@ -282,7 +282,11 @@ class SIMDFastPFor : public IntegerCODEC {
         static_cast<uint32_t>(bc - &bytescontainer[0]);
     *(out++) = bytescontainersize;
     memcpy(out, &bytescontainer[0], bytescontainersize);
+    uint8_t* pad8 = (uint8_t*)out + bytescontainersize;
     out += (bytescontainersize + sizeof(uint32_t) - 1) / sizeof(uint32_t);
+    while (pad8 < (uint8_t*)out)
+        *pad8++ = 0; // clear padding bytes
+
     uint32_t bitmap = 0;
     for (uint32_t k = 2; k <= 32; ++k) {
       if (datatobepacked[k].size() != 0)
@@ -508,7 +512,11 @@ class SIMDSimplePFor : public IntegerCODEC {
         static_cast<uint32_t>(bc - &bytescontainer[0]);
     *(out++) = bytescontainersize;
     memcpy(out, &bytescontainer[0], bytescontainersize);
+    uint8_t* pad8 = (uint8_t*)out + bytescontainersize;
     out += (bytescontainersize + sizeof(uint32_t) - 1) / sizeof(uint32_t);
+    while (pad8 < (uint8_t*)out)
+        *pad8++ = 0; // clear padding bytes
+
     size_t outcap = 0;
     ecoder.encodeArray(datatobepacked.data(), datatobepacked.size(), out,
                        outcap);

diff --git a/headers/simdgroupsimple.h b/headers/simdgroupsimple.h
@@ -630,6 +630,9 @@ class SIMDGroupSimple : public IntegerCODEC {
     __m128i *outDataArea128 = reinterpret_cast<__m128i *>(initOutSelArea8 +
         countSelArea8 + sizeof(uint8_t) + countPadBytes);
     const __m128i *const initOutDataArea128 = outDataArea128;
+    uint8_t* pad8 = (uint8_t*)outDataArea128 - countPadBytes;
+    while (pad8 < (uint8_t*)outDataArea128)
+        *pad8++ = 0; // clear padding bytes
 
     const __m128i *in128 = reinterpret_cast<const __m128i *>(in);